Introduction
Place recognition (PR) is a fundamental component of Simultaneous Localization and Mapping (SLAM), enabling robots to recognize previously visited locations and close loops to correct accumulated drift. While many existing LiDAR PR methods assume a full 360° Field of View (FOV), real-world applications often involve restricted FOVs due to sensor mounting, occlusions, or the use of solid-state LiDARs with narrow scan patterns.
When the FOV is constrained (e.g., to 60° or 90°), traditional descriptors like Scan Context (SC) become sparse and lose descriptive power. Furthermore, high-performing deep-learning methods often require significant GPU resources and large descriptor files, making them unsuitable for lightweight onboard computers or multi-robot systems with limited bandwidth. This paper introduces SOLiD (Spatially Organized and Lightweight Global Descriptor), an ultra-lightweight method designed specifically for FOV-constrained LiDAR place recognition.
Methodology
SOLiD structures and weights spatial data through a novel pipeline that avoids the information loss associated with standard Bird’s Eye View (BEV) projections. The core architecture consists of three main components:
- Spatial Organization (REC & AEC): 3D points are converted into spherical coordinates to create two dense 2D representations. The Radial-Elevational (REC) captures point distribution across distance and height, while the Azimuthal-Elevational (AEC) captures distribution across horizontal angles and height.
- Implicit Elevation Vector (IEV) Reweighting: To handle non-repetitive scan patterns, the system calculates dominant vertical angles (those containing the most points) to reweight the REC and AEC. This emphasizes spatially significant features and preserves vertical structure, which is critical when horizontal context is scarce.
- Two-Part Descriptor: The system generates R-SOLiD for fast loop detection via cosine distance and A-SOLiD for estimating the initial heading (yaw) of the robot.
Experimental Results
The researchers validated SOLiD using the KITTI dataset (clipped to simulate narrow FOVs), the HeLiPR dataset (featuring real solid-state LiDARs), and a multi-robot Park dataset.
Accuracy and Robustness
In 60° FOV tests on KITTI, SOLiD achieved an AUC of 0.987, significantly outperforming traditional methods. As shown in the radar charts below, SOLiD maintains high performance across AUC, F1 max, and Recall@1 even as the FOV narrows, whereas competing methods like Scan Context (SC) degrade rapidly.
Efficiency and Scalability
SOLiD is exceptionally lightweight, with a descriptor size of only 448 bytes—up to 1000x smaller than state-of-the-art methods like RING++. This efficiency translates to massive gains in multi-robot communication; in tests, SOLiD's communication time was just 0.138 seconds, compared to over 100 seconds for heavier descriptors. Additionally, the method runs at 80Hz on a standard CPU, requiring no GPU acceleration.
Conclusion
SOLiD addresses the critical challenge of place recognition for sensors with restricted FOVs. By prioritizing vertical structural information and employing an efficient reweighting scheme, it achieves state-of-the-art accuracy with a minimal memory footprint. While current single-frame descriptors struggle with reverse-direction recognition in narrow FOV scenarios, future work will focus on integrating sequential frame information to overcome this limitation and deploying the method within a distributed SLAM framework.
Summary updated